Side conversations take center stage

Codex enters the ChatGPT app with inline editing and PR review

The biggest shift this week is Codex joining the ChatGPT desktop app on macOS and Windows. Existing users can update and keep their projects, settings, and workflows, and optionally make Codex the default view. The move brings new capabilities: you can edit Markdown and code directly in the app with inline annotations and revision requests, review GitHub pull requests in the sidebar with feedback alongside diffs, and work across multiple repositories in a single project. Computer Use is faster with GPT-5.6, and mobile connection reliability and SSH video rendering are fixed.

Alongside the desktop app release, Codex CLI 0.144.0 adds a writes app-approval mode that lets you declare read-only actions while prompting for writes, and MCP tools can now request authentication interactively without an experimental opt-in. A follow-up 0.144.1 hotfix addresses standalone installs failing when GitHub returns compact or reordered release metadata, and ensures code mode falls back to the embedded runtime when the companion host binary is unavailable. Codex CLI 0.143.0 shipped earlier in the week, enabling remote plugins by default with richer catalog rows and npm marketplace sources, plus system proxy support for macOS and Windows (including PAC and WPAD), Amazon Bedrock GPT-5.6 models, and MCP tool search by default.

Claude Code fixes terminal freeze, auto-mode changes, and security hardening

Claude Code v2.1.207 is the week's most critical release: it fixes a terminal freeze during long streaming responses, remote managed settings being permanently recorded without consent, and a plugin shell-injection vector where user config in commands is now rejected. Auto mode ships without opt-in on Bedrock, Vertex AI, and Foundry, and Claude Opus 4.8 is now the default on those platforms. The v2.1.205 release from earlier in the week adds an auto mode rule that blocks tampering with session transcript files, fixes --json-schema silently producing unstructured output when the schema is invalid, and fixes a critical Windows bug where worktree removal deleted files outside the worktree when an NTFS junction or directory symlink existed inside it. v2.1.203 fixed background sessions becoming permanently unresponsive when the daemon's session token went stale, and worktree isolation bugs where subagents sometimes ran shell commands in the parent checkout. The rest of the week's Claude Code releases include v2.1.206 fixing MCP server timeouts and OAuth re-authentication, v2.1.204 fixing hook event streaming in headless sessions, and v2.1.202 fixing Remote Control command delivery and session naming.

Side chats, conversation search, and agent dashboards

Cursor adds side chats that run alongside your main agent conversation — each side chat is a durable full agent session with context from the main chat, letting you ask questions and explore tangents without interrupting your flow. Conversation search indexes agent transcripts across thousands of conversations via the command palette, and there are redesigned project and repo pickers with scoped search. New cloud agent hooks let you observe prompts, responses, thinking, and subagents.

xAI's Grok Build had a busy week too. Grok Build 0.2.98 pins authentication to API key or OIDC in config.toml, and /context now shows token costs for skills and MCP servers. 0.2.97 adds token usage and cost in headless JSON output, and voice mode is now available for API-key sessions. 0.2.96 fixes agent teardown crashes on slim Linux images lacking the ps command, and authentication retries after token refresh. 0.2.95 lets teams ship default allowed commands via managed_config.toml and fixes session corruption after tool-use history. 0.2.94 opens the Agent Dashboard via /sessions and adds a /goal slash command. 0.2.93 improves MCP permission prompts to show planned arguments before approving, and Esc no longer cancels a running turn. 0.2.92 speeds macOS image paste 65× by reading the pasteboard directly, and user-run shell commands now display their complete output. 0.2.91 keeps voice dictation indicators visible during plan mode review, and 0.2.90 adds a configurable MCP tool output limit and local plugin marketplace sources. 0.2.89 brings voice dictation to Linux and adds a /auto slash command for classifier permission mode.

Devin adds Git-backed blueprints and service user automations

Devin's July 3 release introduces blueprints backed by a Git repository for version-controlling environment configurations, and service users can now create and manage automations via the API. The v3 API supports security scan remediation, and automations can trigger on snapshot build completion with a warning for permission mismatches.